[Mesa-dev] [PATCH 5/6] glsl: Optimize (v.x + v.y) + (v.z + v.w) into dot(v, 1.0).

[Matt Turner]

Cuts five instructions out of SynMark's Gl32VSInstancing benchmark.
---
Could be made more general, but I don't see any cases we could optimize
in our existing shader collection.

 src/glsl/opt_algebraic.cpp | 35 +++++++++++++++++++++++++++++++++++
 1 file changed, 35 insertions(+)

diff --git a/src/glsl/opt_algebraic.cpp b/src/glsl/opt_algebraic.cpp
index eb5544a..17bcdfc 100644
--- a/src/glsl/opt_algebraic.cpp
+++ b/src/glsl/opt_algebraic.cpp
@@ -329,6 +329,41 @@ ir_algebraic_visitor::handle_expression(ir_expression *ir)
       if (op_const[1] && !op_const[0])
 	 reassociate_constant(ir, 1, op_const[1], op_expr[0]);
 
+      /* Recognize (v.x + v.y) + (v.z + v.w) as dot(v, 1.0) */
+      if (options->OptimizeForAOS) {
+         if (op_expr[0] && op_expr[0]->operation == ir_binop_add &&
+             op_expr[1] && op_expr[1]->operation == ir_binop_add) {
+            ir_swizzle *x = op_expr[0]->operands[0]->as_swizzle();
+            ir_swizzle *y = op_expr[0]->operands[1]->as_swizzle();
+            ir_swizzle *z = op_expr[1]->operands[0]->as_swizzle();
+            ir_swizzle *w = op_expr[1]->operands[1]->as_swizzle();
+
+            if (!x || x->mask.num_components != 1 || x->mask.has_duplicates ||
+                !y || y->mask.num_components != 1 || y->mask.has_duplicates ||
+                !z || z->mask.num_components != 1 || z->mask.has_duplicates ||
+                !w || w->mask.num_components != 1 || w->mask.has_duplicates) {
+               break;
+            }
+
+            bool swiz_seen[4] = {false, false, false, false};
+            swiz_seen[x->mask.x] = true;
+            swiz_seen[y->mask.x] = true;
+            swiz_seen[z->mask.x] = true;
+            swiz_seen[w->mask.x] = true;
+
+            if (!swiz_seen[0] || !swiz_seen[1] ||
+                !swiz_seen[2] || !swiz_seen[3]) {
+               break;
+            }
+
+            if (x->val->equals(y->val) &&
+                x->val->equals(z->val) &&
+                x->val->equals(w->val)) {
+               return dot(x->val, new(mem_ctx) ir_constant(1.0f, 4));
+            }
+         }
+      }
+
       /* Replace (-x + y) * a + x and commutative variations with lrp(x, y, a).
        *
        * (-x + y) * a + x
-- 
1.8.3.2